Voice synthesis method based on limited Boltzmann machine

Abstract

The invention discloses a voice synthesis method based on a limited Boltzmann machine. The method comprises the following steps: substituting spectral envelope which is extracted by an adaptive weighted spectrum interpolation STRAIGHT synthesizer for high-level spectrum feature for spectrum modeling; performing state segmentation on acoustic feature sequence in a training database by using a Gaussian-hidden Markov model (HMM) model obtained through training; segmenting the original spectral envelope feature of the extracted training database by utilizing the starting and ending time of each state obtained through segmentation, and acquiring spectral envelope data corresponding to each state in a context related HMM model; and predicting the spectrum feature by using the Gaussian-HMM mode, feeding the spectral envelope feature obtained through prediction and base frequency feature into the STRAIGHT synthesizer and generating the final synthesized voice. By the method, the spectrum feature modeling precision of an HMM-based parameter voice synthesis method can be increased, so that the tone quality and the naturalness of the synthesized voice can be improved.

Description

technical field [0001] The invention relates to the technical field of signal processing, in particular to a speech synthesis method based on a restricted Boltzmann machine. Background technique [0002] At present, speech synthesis realizes the conversion of text to speech, and is one of the core technologies of intelligent human-computer interaction. Parametric speech synthesis based on Hidden Markov Model (HMM) is a mainstream speech synthesis method at this stage. During training, the method first extracts the acoustic features such as frequency spectrum and fundamental frequency in the training speech database, and then uses a unified HMM framework to model the acoustic features; when synthesizing, firstly use the statistical model obtained from the training based on the maximum output probability criterion. The prediction of various acoustic features, and then the predicted acoustic features are sent to the parameter synthesizer to reconstruct the synthesized speech. ...

AI Technical Summary

Problems solved by technology

However, the sound quality of synthetic speech is often not ideal, resulting in poor overall naturalness

[0003] The insufficiency of the above-mentioned traditional speech synthesis method based on HMM parameters in spectrum modeling is an important reason for the unsatisfactory sound quality of synthesized speech

Specifically, since the spectral features used in traditional spectral modeling are often some high-level spectral features, such as Mel Cepstra (Mel Cepstra), Line Spectral Pairs, etc., these features are based on the original speech A modeling or approximate representation of the spectrum has caused the loss of spectral detail information in the process of feature extraction; at the same time, since the traditional spectral modeling method usually uses a single Gaussian distribution to describe the spectral feature output of each state in the HMM Probability, in the synthesis stage, the spectral features are predicted based on the maximum output probability criterion. Since the mean value of the single Gaussian distribution has the largest output probability, the parameter generation result is very close to the mean value of the model, and the mean value is based on the maximum likelihood in the training phase. The criterion is estimated by averaging the training samples, which causes the predicted spectral features to be too smooth, thus affecting the sound quality of the final synthesized speech

Images